The Knowledge Graph Divide - Connecting Machine Learning, Databases, and the Semantic Web

Abstract

Over the past decade, Knowledge Graphs (KGs) have received enormous interest from industry and academia. However, there are three key research communities, namely the Machine Learning (ML), Database (DB), and Semantic Web (SW) communities, studying KGs with major gaps between them. This dissertation is about bridging their divisions: Reasoning Divide. KGs are inherently incomplete. Therefore, the ML community has proposed Knowledge Graph Embedding Models (KGEs), achieving promising results for predicting missing links. Key data properties in the DB and SW fields are typically represented via logical rules. However, any current KGE cannot capture vital rules, i.e., infer missing links while adhering to such rules. Capturing (i) general composition and (ii) composition and hierarchy rules jointly are crucial open problems. To bridge this division, we introduce the ExpressivE model that embeds pairs of entities as points and relations as hyper-parallelograms in the virtual triple space R2d. This model design allows ExpressivE to capture a rich set of logical rules while offering an intuitive and consistent geometric interpretation of ExpressivE embeddings and their captured rules. Scalability Divide. Even more, the SW and DB communities provide massive KGs, calling for efficient KGEs. However, most contemporary ML-based KGEs require high-dimensional embeddings or complex embedding spaces for competitive prediction results, drastically raising their space and time requirements. Thus, developing efficient KGEs makes up another central open problem dividing the SW, DB, and ML fields. Facing this challenge, we propose SpeedE, a Euclidean KGE that (i) has strong inference capabilities,(ii) is competitive with state-of-the-art KGEs, significantly outperforming them on the YAGO3-10 and WN18RR benchmarks, and (iii) dramatically increases their efficiency, needing on WN18RR solely a fifth of the training time and a fourth of the parameters of the best-performing model (ExpressivE) to reach the same link prediction performance. Data Management Divide. Above all, the DB and SW communities have driven classical KG research. However, there remains a divide between approaches from these two fields. For instance, while languages such as SQL or Datalog are widely used in the DB area, a vastly different set of languages, such as SPARQL and OWL, is used in the SW area. This mismatch, however, makes blending KGs from both communities a complex endeavor, rendering the interoperability between DB and SW technologies a pressing open challenge. Thus, we present the SparqLog system, a uniform and consistent KG management framework meeting essential requirements from the SW and DB fields.